AI investment logic has shifted: Cloud providers enter the "landlord" era, infrastructure supply chain under pressure

Original Author: Long Yue

Original Source: Wallstreetcn

In the same earnings season, Microsoft's market cap surged by $450 billion in a single day, setting a record for the largest single-day market cap increase in global stock market history, while Amazon crossed the $3 trillion mark within five days. On the other hand, SK Hynix's stock price was halved, Kioxia's stock price was halved, and nearly all global memory stocks fell over 40%.

Both are doing AI. Half are surging, half are being cut in half.

This may be the market using real money to re-answer a core question: In this AI feast, who is actually making money, and who is footing the bill?

Over the past two years, the market rewarded "whoever spends the most" — every time a CSP (Cloud Service Provider) raised its capex guidance, GPUs, HBM, optical modules, switches, PCBs, power supplies, liquid cooling, and data centers all rallied in tandem. Now, the market is starting to ask a stricter question: Of the trillions of dollars that tech giants are pouring in, whose cash flows are they flowing into, and whose depreciation are they settling into?

The answer may be emerging — AI money hasn't disappeared; it has just changed pockets. It flows from those who sell the shovels to those who turn shovels into businesses. It flows from those who build the roads to those who collect tolls on them.

Three signals flash simultaneously, and the market begins re-pricing

To understand this round of divergence, one cannot just look at the rise and fall numbers. In this time window at the end of July, three things happened at once.

First: CSPs (Cloud Service Providers) proved they can turn computing power into revenue.

Microsoft Azure grew 43% year-over-year. Amazon AWS grew 37% year-over-year, the fastest in 18 quarters. Google Cloud grew 82% year-over-year, the highest in three years. Together, the three generate $389 billion in annualized cloud revenue, with a quarterly net increase of $50 billion in annualized recurring revenue — double the average of the past three quarters.

More importantly, revenue is accelerating, but not just because a few AI labs are buying computing power. Microsoft disclosed a key figure: commercial remaining performance obligations reached $678 billion, up 84% year-over-year, and all of the sequential growth came from non-frontier lab customers. Banks, manufacturing, healthcare, government — those traditional enterprises that the market believed "AI has nothing to do with them" over the past two years are now signing contracts.

AWS isn't just talking — AWS is doing. On August 4, AWS head Matt Garman publicly confirmed that the company is signing five-year commitment contracts with customers, adding that "demand continues to outstrip supply, and we're working hard to accelerate our construction and investment pace to keep up with customer demand." A five-year contract means customers have embedded it into their core business processes — this kind of revenue stickiness is the most powerful antidote to any "capex anxiety."

Second: The narrative model for the infrastructure chain has failed.

SK Hynix posted its strongest-ever revenue and profit — operating profit surged 557%. Then the stock price plummeted.

Why? When actual growth can't outpace forward expectations, and when CSP capex year-over-year growth slides from double-digit acceleration to high single digits, the valuation model has to change — from "how much can you sell" to "how fast can you still grow."

This is a law of physics, not a judgment call. CSP capex cannot permanently maintain 30% or 50% year-over-year growth. When the industry moves from an "accelerating expansion phase" to a "high-level growth phase," the second derivative of growth turns negative, and valuations must be re-priced. That's why Micron, Kioxia, and SK Hynix experienced their most brutal selloffs while reporting their strongest financials.

Third: Model pricing power is collapsing.

In the busiest week of earnings, OpenAI cut the price of its new model GPT-5.6 Luna by 80%. From $1 per million tokens, straight down to $0.2. It had only been live for three weeks.

This is a price cut forced by DeepSeek. Over the past year, Chinese open-source models' share of token calls on the global API platform OpenRouter surged from 4.5% to 46%. DeepSeek V4 Flash is priced at $0.14, and Alibaba's Qwen3.7 Flash is even lower at $0.03. The toll booths on model access are being dismantled.

Put these three things into the same time coordinate —

CSPs can make money + infrastructure growth peaks + model pricing power collapses = Profits are flowing back on a large scale from the hardware and model layers to the infrastructure and platform layers.

Why can CSPs collect tolls? The cloud giants' ROI ledger: Heavy assets ≠ Low returns

To understand why CSPs have suddenly become the market's new favorite, you can't rely on intuition — you need to run the numbers.

First calculation: The profit ledger — The cloud giants' ROI ledger, heavy assets ≠ low returns

Over the past two years, the market assumed cloud giants were the ones getting taken advantage of — all the money was being made by Nvidia and Hynix. Morgan Stanley's estimates at the end of July overturned this judgment: Pure GPU leasing benchmark ROIC (Return on Invested Capital) is approximately 31%. Under the assumptions of a 1GW data center, 410,000 GB300 chips, and 75% utilization, if leasing prices rise from $7 to $10 per hour, ROIC can jump from 23% to 39%.

And don't forget — the core of a data center — land, power, and facilities — has a lifespan of over 30 years, spanning five to six generations of servers. Every server generation updates, but the infrastructure doesn't need to be rebuilt. This means the CSPs that entered first will have better marginal economics in subsequent generations, not worse. The market eventually realized this.

Morgan Stanley broke down bottom-up ROIC estimates for three AI business models:

Pure GPU leasing (IaaS): Benchmark ROIC of approximately 31%. Under the assumption of GPU supply shortage and 75% utilization, if leasing prices rise from $7 to $10 per hour, ROIC can jump from 23% to 39%.

Model APIs on proprietary infrastructure: Benchmark ROIC of approximately 46%. This assumes model companies maintain product differentiation and pricing power — but given the current price war, this assumption is wavering.

Model APIs using third-party computing power: ROIC of only 25%. They need to pay an additional "middleman margin" to cloud providers.

The core conclusion of this analysis directly challenges a prevailing assumption the market has held for nearly a year: Heavy assets ≠ Low returns.

Second calculation: The revenue ledger — Demand has spread from labs to enterprises

Over the past two years, the market's biggest fear about AI capex was "circular financing": CSPs pour money into model companies like OpenAI and Anthropic, and then model companies buy back CSP computing power. Passing money from the left hand to the right — where is the real end-user demand?

This fear is being disproven by this quarter's data.

Microsoft's commercial remaining performance obligations saw all sequential growth come from non-lab customers. AWS's backlog has reached $496 billion, with triple-digit growth. Google Cloud's backlog is $514 billion, increasing by $50 billion in a single quarter. Nearly 90% of Fortune 100 companies are already using Gemini Enterprise.

Wallstreetcn previously used a vivid metaphor: CSPs were previously building a house for a single tenant, the "frontier lab." Now the entire office building is starting to fill up.

This isn't an isolated signal from one company. All three are showing the same trend simultaneously — AI demand is spreading from a small number of startup AI labs to enterprise customers in traditional industries. When the customer base shifts from "concentrated" to "broad," the circular financing narrative no longer holds.

Third calculation: The model ledger — Four companies, four ways to collect tolls

CSP is not a monolithic concept. What's truly interesting about this earnings season is that the four giants have formed four completely different paths to AI returns.

Microsoft — A three-tier toll collection model.

Azure provides the most basic GPU leasing and cloud services — that's tier one. The Foundry platform offers model invocation, fine-tuning, and deployment — that's tier two. Copilot is embedded directly into Office 365 and GitHub, charged per seat (paid seats have exceeded 30 million) — that's tier three.

The three tiers funnel traffic to each other: Azure enterprise customers can upgrade to Foundry to use models; models on Foundry can penetrate Copilot's application scenarios; and Copilot in turn drives Azure consumption. This isn't a revenue line — it's a revenue matrix.

Amazon — The simplest and most transparent toll collection model.

AWS itself is a business that rents out computing power and storage. It has been operating for 18 years with an extremely mature customer, billing, and retention system. This quarter's revenue was $42.2 billion, up 37% year-over-year, the fastest in 18 quarters.

Google — The flashiest numbers, the fuzziest ledger.

Google Cloud's 82% growth, profit margin, and margin improvement appear to be the strongest of the four on the surface. But the market's response was to fall first and then rise.

The reason lies in the "gray zone" of cost attribution. Gemini simultaneously serves Search, Ads, Workspace, YouTube, consumer apps, and cloud services, with shared AI R&D expenses remaining at the group level. Google Cloud's segment profit margin cannot be equated with Google's true economic margin across its "chip-model-cloud" full stack.

Some research institutions have pointed out that if all Gemini expenses were charged back, Google Cloud's profit margin would only be around 10% plus. Google isn't not making money — it's just that the path to profitability is too complex.

Meta — Going from "pure spending" to "starting to think about collecting tolls."

Over the past two years, Meta has been the most awkward of the four in investors' eyes: AI has indeed improved ad recommendations, content ranking, and creative generation, but these benefits are hidden within the existing advertising business and cannot be matched line-by-line to new capex. When capex guidance was raised from $130 billion to $145 billion, leaving free cash flow at just $784 million (down 91% year-over-year), the market's patience reached its limit.

But a subtle shift is happening at Meta.

The company has been revealed to be planning a cloud business, considering offering model access services to developers and selling surplus AI computing power. This means Meta's computing assets are no longer a binary choice between "self-use or idle." It can switch to external leasing.

The model price war: The biggest tailwind for toll collectors

The most easily overlooked trigger variable in this round of divergence isn't earnings — it's the price collapse at the model layer.

Go back to July 30. Besides Microsoft's earnings, OpenAI did one other thing: it cut the price of GPT-5.6 Luna to one-fifth of its original. Not a price cut after a year on the market, but a price cut three weeks after launch.

Behind this is a larger trend — Open-source models are dismantling the "toll booths" of closed-source models.

Over the past two years, the profit distribution logic of the AI industry chain has been: Nvidia earns money from GPUs, CSPs earn the hard-earned money from computing leasing, and model companies (OpenAI, Anthropic) earn premiums through their technical moats. The model layer is a "toll booth sitting on top of the cloud" — no matter whose computing power you use or what application you run, calling the strongest model requires passing through it.

Now that toll booth is being dismantled.

The breakthrough of Chinese open-source models is the biggest sledgehammer. DeepSeek V4 trained a 1.6 trillion parameter model using Huawei Ascend 950 chips, with performance comparable to GPT-5. Alibaba's Qwen3.7 Flash is priced as low as $0.03 per million tokens for input. On the largest model API aggregation platform, OpenRouter, Chinese models' share of token calls surged from 4.5% to 46% in one year.

When 46% of calls go through cheaper open-source models, closed-source models can no longer charge a "we're the only one" premium. OpenAI was forced to follow with price cuts — not because it wanted to, but because it had no choice.

What does this mean for CSPs?

Barclays analyst Raimo Lenschow wrote a remarkably precise assessment in his August 3 report: When models go from "scarce goods" to "commodities," pricing power flows back to the platforms that own computing power, customers, and distribution channels.

CSPs no longer have to be tied to any single model company. They can simultaneously offer GPU leasing + self-hosted open-source models + proprietary model APIs. Customers don't have to choose just one model; CSPs can act as "routers" — high-complexity tasks use top-tier closed-source models, routine tasks use cheaper open-source models, and high-throughput, low-cost steady-state workloads run on proprietary ASIC chips.

Who ultimately benefits most from this "multi-model architecture"? Not the company that's best at training models, but the company that's best at chaining different models together, allocating computing power on demand, and charging by usage. That's the CSP.

Morgan Stanley's July 27 report even broke this logic down into math: If model API prices drop 50%, but token call volume consequently rises 5x, it's a net positive for inference revenue on CSPs' proprietary infrastructure. Because the cost side is following the efficiency curve of chip performance and software optimization, while the revenue side is following the usage curve of exploding demand. As long as the second curve runs faster than the first, it's a good business.

For independent model companies, a 50% price drop means revenue is cut in half — unless token call volume can double, but the substitution effect of open-source models makes that impossible.

Who's footing the bill: The brutal mean reversion of the infrastructure chain

If CSPs rising is "the toll-collection story being priced in," then the infrastructure chain's crash is "the growth myth being exposed."

Morgan Stanley strategist Michael Wilson used a precise analogy: Semiconductor stocks, especially memory, trade remarkably similar to silver.

There are two strands to this logic. First, both experienced parabolic rises — sentiment and capital pushed prices to levels that fundamentals couldn't support. Second, both have commodity attributes —